Sie sind auf Seite 1von 7

Performance Enhancement of Downlink Multiuser

DS-CDMA Detectors Using Processing by

Independent Component Analysis


Said E. El-Khamy, Fellow IEEE, and Mansour Abed
Department of Electrical Engineering, Faculty of Engineering, Alexandria University, Alexandria, Egypt
Email: said.elkhamyggmail.com, man.abedgyahoo.fr

Abstract- In this paper, the attention is focused on the such details as the number of co-existing users served by the
generation of the received digital signal and its linear detection same uplink (i.e. base station) at any instant of time, the codes
over a synchronous multiuser direct sequence code division of those interfering users and the multipath channel delays of
multiple access (DS-CDMA) downlink system considering
various spreading codes, additive white Gaussian noise (AWGN) the received signal are known by the downlink. Still, the goal
and introducing varying signal-to-noise ratios (SNRs) and users' is to detect the transmitted data symbols from the received
power distributions. The aim of the work is to test the efficiency signal and so the concept of blind MUD techniques [4, 5] is
of implementing an independent component analysis (ICA) based introduced for better estimations of the user signal and easing
estimator of the transmitted symbols preceded by principal the computational load.
component analysis (PCA) pre-processing of the received data to The researches on ICA-CDMA framework were motivated
reduce dimension and decrease the noise effects. The noisy ICA by the fact that the data received through slowly fading
algorithm is used either as: (1) a pre-processor prior to
conventional detection: The idea here is to estimate the mixing multipath channel DS-CDMA downlink can be modeled from
matrix that contains the basic vectors and fading terms using the ICA viewpoint by simply transforming the blind multiuser
separating matrix and PCA subspace or as (2) a post-processor detection (BMUD) problem to an equivalent instantaneous
attached to a linear DS-CDMA receiver. Numerical simulations blind source separation problem. Joutsensalo and Ristaniemi
indicate that the first schema could constitute a powerful tool for were the first to apply ICA learning on CDMA linear
channel and symbols estimations and has the advantage that it detection. In [6], the authors compared two different iterative
doesn't necessitate any prior knowledge of the users' codes or learning approaches in the CDMA downlink symbol
paths delays and strengths except providing a short pilot
sequence (presenting 1% to 4% of the transmitted symbols of the estimation: The first is a noiseless fast fixed-point algorithm
user of interest) that is used to give a good enough initial guess of and the second is a semi-blind iterative algorithm that assumes
the demixing matrix columns and so force the ICA iteration to be a binary form of the source data. In [7], a novel scheme for
in the wanted user subspace. The second schema shows the separation is considered by extending the matched filter
performance comparable to or superior than the exact blind technique by the means of independent component analysis.
detectors at relatively high SNR values. The same authors proposed in [8] a FastICA based separation
scheme that utilizes the known code of the user of interest.
Index Terms- Blind source separation, independent component The scheme needs the estimation of the p strongest path
analysis, multiuser detection, principal component analysis
delays. The application of the noise-free ICA principle in
downlink CDMA timing acquisition was analyzed in [9, 10].
I. INTRODUCTION In [11], ICA for complex valued mixtures was considered as a
Code division multiple access (CDMA) is a spread tuning element, which was attached to an existing DS-CDMA
receiver structure. The researchers applied in [12] the natural
spectrum technique that is intrinsically impaired by multiple gradient based online blind signal recovery (NGBSR)
access interference [ 1 ]. In direct sequence CDMA techniques using either the feedforword or the feedback
communication systems, the narrowband signal energy of each symbol recovery structures to achieve the global task of
user is spread over a wide frequency range using pseudo-noise
BMUD for QPSK DS-CDMA methods assuming equal users'
(PN) sequences that are independent of data so that the effect power distributions. A noisy ICA based forward CDMA
of fading channels is reduced. However, simultaneous use of receiver for multiple access communications channels was
the same frequencies allows different users to interfere with proposed in [13]. The algorithm is used as post-processor
each others which limits the system capacity unless taken care attached to a subspace based CDMA receiver in the presence
of. Multiuser detection (MUD) [2] is a technique that includes of Gaussian noise and considering equal power distribution.
all the users in the signal model and exploits the interference Three adaptive natural gradient based user detection
structure to improve data recovery. However, the optimal
M\UD [3] is computationally exhausting and requires many approaches applied for both DS-CDMA and wideband CDMA
system parameters to be known. In fact, the key problem in the
(WCDMA) systems have been published in [14].
In this paper, independent component analysis (ICA)
CDMA downlink (i.e. mobile phone) signal processing is that [15, 16] principle is exploited to achieve blind channel
downlink has no prior information except its own code. No

1-4244-0272-7/06/$20.00 ©2006 IEEE 237


identification using the separating matrix and principal III. DS-CDMA CHANNEL MODEL
component analysis (PCA) subspace. Then, detection of the
desired user's symbols is performed using reference methods In this section, a downlink CDMA channel model is
namely blind MMSE and blind decorrelating detectors. In the described and shown to be similar to a noisy ICA model. The
other hand, enhancement of downlink detection using an ICA signal studied in this paper is a block slowly fading, multipath,
post-processor attached to an exact blind detector is synchronous DS-CDMA downlink AWGN model. The data in
investigated and evaluated. The comparative tests include long the observation interval have the form [6-8]
N K L
spreading codes, large network sizes and AWGN effects r(t) = b,Ykn Y,alSk (t -mT -dl) + n(t) (4)
which are neglected in most of the works done in ICA-CDMA m=1 k=1 /1=
applications. Also, the influence of power dissimilarities under Here, N symbols are sent by K users and received via L paths,
severe selective fading multipath channels is considered and al is the attenuation factor of the fIt transmission path, also
discussed in detail. called path gain, containing the corresponding path power
strength as well, bkm is kth user's mt" symbol, Sk(.) is kth user's
II. NoisY ICA MODEL FORMULATION binary chip sequence, Sk(t) E {- 1,1 }, of length C and
supported by [0,1), where T is the symbol duration, the
ICA is a higher statistical technique that is mainly used in discretized delay of the 1th path is denoted by dl, dl E {0,
blind source separation (BSS) applications. Its goal is to (C-1)/2}, and it is assumed to change sufficiently slowly for
express a set of random variables as a linear combination of most of the time, and n(t) denotes noise. The processing gain
underlying latent variables [17]. In the basic form of ICA, we in a direct sequence spreading is equal to the chip sequence
observe n scalar random variables r1,r2,...,rn which are linear length C.
combinations of I unknown independent components, or ICs, As the continuous-time data is received, it is first sampled
denoted by bj,b2,...,b,. The ICs are assumed to be mutually by chip-matched filtering. In our case we use chip rate
statistically independent, non-Gaussian (except for one) and sampling, i.e. C equispaced samples per symbol are taken. The
we must have at least as many observed linear mixtures as samples are then collected into C-vectors rm,
ICs; i.e., n> 1. Let us arrange the observed variables ri into a
vector r=(rj.r2,...,rn)T and the IC variables bj into a vector rm=[ r[mC] r[mC+1] ... r[(m+1)C-1] ]T (5)
b=(b1.b2,...,b,)T. In the literature, one popular way of Since sampling is generally symbol-asynchronous, the vector
formulating the linear ICA (noise-free) problem is to consider sampler rm usually contains information about two successive
the following generative statistical model for the data [15, 17- symbols, and the data thus have the well known form [20]
19]
K L
r=Gb (1) rm =E b a, gk + b, E al g, + nn (6)
k=l 1=I /=I1
In the noisy ICA model, the noise is assumed to be additive
and the observed data can be expressed as Here nm denotes the noise vector, and the "early" and "late"
rm=Gbm+nm (2) parts of the code vectors are respectively
where rm is the mth observed data vector, G is an unknown full gkl =[0k[Cd. +11] ... Sk [C] 0 d.] 0] (7)
rank mixing matrix, bm is an unknown non-Gaussian source
process and nm is an additive Gaussian noise process. The goal (8)
is to estimate the noise-free ICs bm using only the observations gkl =
O ... ° Sk [1] "
Sk [C-d, ]H
rm and the assumption of the independence of the sources. This Equation (6) can be represented more compactly as
is roughly equivalent to estimating the mixing matrix G [18].
In most references, however, the source estimation means that rm=Gbm+nm (9)
a set of filters w1, w2, ... should be estimated such that
W= [w1, w2, ...] is the separating or decoupling matrix; then, where the Cx2K dimensional matrix G -[g1,.., g2K] is
the output source estimations w1Trm, w2Trm, ... or equivalently assumed to be full rank such that,
ym= WTrm (3) L L L L

are as independent as possible and each of them can be used to


G= Yag-1/ 'ya,
-/=I
,

/=I
,
911'...,Ya1gK1lYal9K1
/=I /=I
(10)
represent one of the latent sources. The quality of the source
and the 2K-vector bm contains the symbols,
estimates, or in other words the success of the implemented
separator to recover the latent random variables, can be
measured using the combined system matrix defined, for non- bm=[bi,mi,, bim, ..., bKm-1, bK,m] (1 1)
whitened data, as: P=WG, that should approximate as closely
as possible a scaled permutation matrix, i.e.: P-4DD where TD The similarity of the model (9) to noisy ICA model (2) is now
is a permutation matrix with one unity entry in any row or immediately seen. In (9) we also have an unknown mixing
column and D is a diagonal nonsingular scaling matrix. matrix (here G) and an independent (or roughly independent)

238
source process (here bin). Our goal is to recover the syrmbol E{Z ZTI =I1+±(As- 2 I-' (19)
vector bm for each user k directly without knowing the c aodes
or parametric form of G. In the pure matrix form Thus, the noisy ICA learning rule must take into account the
representation of the data is non-whiteness of the data. This conducts us to the introduction
of the well known FastICA algorithm for noisy data as
X=GB+N (12) described in the following section.
where B. FastICA Algorithmfor Noisy Observation Data
X=[rl,...,rN] (13) Hyvarinen and Oja's FastICA fixed-point batch algorithm
[15] is a contrast-based approach where the independent
B=[b1,...,bN] (14) component analysis data model (1) is estimated by
N=[n1,. . . ,nN] (15) formulating an objective function (for W) to be optimized
under some assumptions and constraints. Such a function
In (12), G depends on the channel coefficients al, delays dl and serves as a measure of independence of the output solutions,
the codes Sk(.), while the 2KxN matrix B depends only on the and is often called contrast function. If noisy observations are
symbols bk,. considered, the ICA problem (12) can be solved using a
modification of the FastICA algorithm as introduced in [22].
When kurtosis is used as the contrast function, the following
IV. ESTIMATION ALGORITHM one-unit learning rule is formulated
A. PCA Pre-processing 1. Quasi-whitened the noisy data using (17)
2. Take a random initial w1=wo
ICA is usually preceded by a whitening transform which
makes the components of the observed mixtures mutually 3 . Iterate w1 = -3 (I +_) w, w, (I + Z) wl,
uncorrelated and normalizes their variances to unity.
Simultaneously, the dimensionality of the data is reduced and where X =& (As - a I)1
some additive noise is suppressed out. This can be done for
the noiseless case by PCA as follows 4. Divide w by its norm: w* = w |
Y =A -'2UT X1f =A -112uT GB (16) 5. If sign(ZTw7) = sign(ZTwl) then stop the iteration, else
update w1 = w7 and go to step 3
where As represents the 2K-dim matrix of principle
eigenvalues of the data correlation matrix estimate Mc= Note that in step 3, (.)3 means elementwise operation.
XXT/N and As-,12 is computed by simple component-wise Since in the CDMA downlink only one signal is of interest,
operation: As 2=diag[,A21/2 ..,21/2], Us denotes the 2KxC it would be desirable to remain with estimating only one basic
vector as described in the above procedure. However, it is
matrix of corresponding principal eigenvectors of the data extremely important to give good enough initial guess for wo
correlation matrix Mc, and T represents the Hermitian in order to ensure that the estimated column of the separating
Transpose operator. matrix W is that corresponding to the user of interest.
Consequently, applying PCA reduces the signal space to a Furthermore, if we want to recover the whole ICA system; a
dimension in which two sources correspond to each user's decorrelation step must be introduced in order to prevent
main transmission path. In fact, by specifying a convenient different neurons outputs wjTZ,..., WTKZ from converging to the
threshold for the eigenvalues to be retained from the PCA
decomposition, we can select only those principal components same maxima. A simple way of achieving decorrelation is a
that explain well the observed data and simultaneously, this deflation schema as explained below
may constitute a tool for estimating the number of users K. Let p, 1<p<2K, be the index of some independent component
This well known PCA approach can be extended to whitening to be estimated, then the following orthogonalization must be
of noisy data using bias removal technique [21], however the introduced in the beginning of step 4
T
resulting filtered variable covariance matrix is not unity and so w
p= wp - wp wp-l wp-l , (20)
the term of quasi-whitening is utilized. = W(1, p -1,1, 2K); w = W(p, 1, 2K), p = 2, 3,...,2K
In fact, quasi-whitening operation can be performed on the
noisy data as follows After the separating matrix estimate has been found, the
Z =TX= TGB+N (17) transmitted symbols of the user of interest can be calculated
where N is the linear transformation of the Gaussian noise and using the following decision formula
T =(A- 2 i)-1/2 u1T (18)
bp = sign{(wp)TZ}, p=1,2,...,2K (21)
The 2KxN matrix Z presents the quasi-whitened data and Or, for multi-unit learning rule, the 2KxN matrix B estimate
follows a noisy ICA model. The covariance matrix of Z is is given by
different from the identity matrix since [13] B = sign{WT Z} (22)

239
C. Flowchart of the Algorithm It is important to note that the FIR channel coefficients
calculated above must be adjusted according to the considered
The purpose of the implemented simulations is to approach power distribution so that
from realistic systems behaviors. The fading multipath
channel is modeled as an FIR filter where the taps are al,k power dist(k). al,k 1 Il,..,L; k-l,.. ,K (29)
normally distributed and the delays are uniformly distributed The pilot sequence length used for separating matrix initiation
within a defined interval. This statistical characterization is denoted by Ps. Fig. 1 presents the flowchart of the
defines what's called Rayleight fading which is the most
common model used in radio communication propagation algorithm.
channels. Moreover, power distribution effects are analyzed. Input Parameters Start
C, code type, K, L, N, As
Here, it is desired to test the efficiency of the investigated Us, X, Ps
receivers if the target user is set to a power lower than the B1=B(s,m); Input Parameters
other active users due to fading, caused by either multipath { s =1:2K; m= 1:Ps } +
propagation or terrain configurations, that degrades the
strength of the transmitted signal to the user of interest.
SNRdB, power dist T=(A --I)
-1/2 U T Z=TX
Z =Z(s,m); s= 1,..,2K; m= 1,..,Ps}
The multipath channel is constructed as follows
Given the maximum delay spread: max delay_spread=(C- Wo=ZI (BA),a = I0(ISNRdB /10)

1)/2, the number of users K, the number of paths L and the 5,=Sa(As -a20
standard deviation of the multipath amplitudes sigmafading,
the paths's delays are expressed by
dlk = rand. mcx delay spread ] 1= 1, .. ,L; k= 1,.., K
_ (23)
where rand is a uniformly distributed random number in the w,=W,,(s,k), s = 1: 2K|
interval (0,1) and the symbol F 1 represents the ceiling .i
operation. Thus, for each given k, L different delays in the wT =-Z ZTw - 3 (I +X) w1wT (I +X)
range {0, ..., (C-1)/2} are generated.
Wnew=;;wvt(s,k)= w , s =1: 2K
The multipath gains are given by
a/k a(d1,k,k)=sigmafading. randn 1=1,..,L; k=l,..,K (24) Yes >~No
and randn is a normally distributed random variable with zero
mean and unit variance.
According to the following input parameters, the power Wprev=wout(S, ,t), s=1:2K; t=1:(K-1)
distribution (denoted by the vector power_dist) is easily (wV*)T (V* )T Wprev (Wprev)T; Wnew= (Wne )T
designed
distfactor: It has the following possible values
1: Log-normal power distribution;
2: Equal power distribution and,
3: Customized power distribution by defining a specific Yes
user's power expressed in dB and denoted by the parameter
custompower,
custom user: Index referring to the user for which power
customization is intended,
sigmapower: Standard deviation of users' power distribution. L;
k
ws,
k
(WV)I
= Sb
^

=-VVse Z;kI
k
Then, set Sb Sb powerdist(k); bk sign(gb)
power_dist(k)= k-=1 K .., (25)
If distfactor= l then kfk+I
power dist(k)= 0(sipapower randn) k= l.. K (26) No Yes
Ceck if. k<2K
If distfactor-2 then preserve equal power distribution for all
the system, and if distfactor-3 set I
power dist(custom user)= 1 0(custompowerl1O) (27) §B=l b Sb Sb0 Bb 2 b2KX
Finally, the kth user's power distribution is normalized W [wL~wlep ep...ws2p 0]T; G=T WT; P=WTG
top

power _ dist(k)=power _ dist(k) power dist(k)jJ (28)


Figure 1. Flowchart for downlink DS-CDMA channel estimation and symbols
detection based on noisy FastICA algorithm and PCA pre-processing.

240
Remark: T denotes the pseudoinverse of the matrix T and TABLE I
satisfies: TT T= Tand T T *=T. PARAMETERS SELECTION FOR A DOWNLINK MULTIUSER DS-CDMA SYSTEM
Introducing blind MMSE and blind zero-forcing post UNDER SLOW SELECTIVE FADING AND EQUAL POWER DISTRIBUTION
detectors provides the (2KxN) recovered data matrices Parameter Value Parameter Value
PN codes type Gold # of users 10
denoted by BNiyCAISEand
BAIYC AVVS BNosIAZ
oiyC F that are formulated
Code length 31 Delays range (in chips) {0,..,15}
respectively as follows Power distribution type Equal # of multipath rays 7
Number of simulated 1000 Std. deviations of multipath 0.4; 0.5
For each k; kl1,..,2K, the N-dim row vectors bNoisyICA-AMSE and symbols and users' power dist.
Pilot sequence length 13 SNR (dB) 10
bNoisyICA-ZF are given by
Figs. 2 and 3 show, respectively, the ICA-based estimate and
bNoisyIcA-AMSE = sign(gj Us As s Xm ) the exact mixing matrix columns related to user 8. Both
I3km =
impulse and frequency responses are pictured. The figures
bNoisyICA-ZF= sign(gj Us (As - aI)1 Us Xm)
(30) indicate that even with a low SNR value and a severe selective
Xl =G(c,k); c= 1,C N
fading multipath, the noisy ICA algorithm estimates
Xmn = X(c,m); c = l, C, m = 1, N efficiently the code-fading system matrix G.
Thus 8's Esfimaf4 :omp6hehfs

NoisyICA-AIEISE . NoisyICA-AIEISE

b2,1 b2,N 10 5 10 15
Chip Duration T.=TIC
20 25
Noisy NoisyICAESE . NoisyICA. .ASE FrequencV Response olF User B's lEstimdted 1Eae1
10 0 0
20

b2K,1
NoisyICA-AIEISE * b2K,N
NoisyICA-AIEMSE
i -10
LE -20 -08 -06 -0A4 -02 0 0,2 O4 06 0~8
Sampling Frequency f 1/TC
NoisyICA -ZF * NoisyICA -ZF WO B's Estimafed Lafe Part drnponehts

b2,1 b2N
BNoisyICA=ZF NoisyICA-ZF . NoisyICA-ZF
-2
0 10 15 20 30
Chip buration T1=TIC
b2K,1
NoisyICA-ZF . b2K'N
NoisyICA-ZF (31) 2 20_ ,ncy Response of User 8's Estinmnated L

Note that the recovered symbols when the ICA algorithm is a


& 10
a10
used as a post-processor are obtained with the same way as 2! -10 _
explained in the flowchart of Fig. 1. The only change is in the 08 06 -0A4 =02 0 02
Sampling Frequency f, 1/T;
04 O 08

separating matrix initiation step so that:


Figure 2. Time and frequency characteristics of user 8's estimated mixing
Wo Z (BBlindASE)T; BBfindMMSE = sign(GT Us As UT X) (32) matrix columns.
S's Exact Early Part Component
for blind MMSE/Noisy ICA detector and similarly
0.5
Wo Z (BBIind ZF)T; 1BIindZF sign(GTUs (A, _I) UT X) (33) -0.5(,
0 5 10 15 20 25 30
for blind ZF/Noisy ICA detector. The enhanced recovered Chip buration
Frequency Ptspnse User B's
of
T.
ixact Garly Part
symbol matrices at the output of the noisy FastICA algorithm a 10

are denoted by BAMSE-NOiSYICA and B ZF-NoisyICA respectively. t 101


o

2, -i 0 -08 -06 -04 -02 0 02 04


0~6 0.8l
Sampling Frequency f.
Urrr E oo
V. RESULTS AND COMMENTS 1 1 s

In what follows, some standard measures that serve to


compare between the investigated detectors are presented by -2L
0 10 15 20 25
referring to the detectors impulse and frequency responses and Chip buration T
the achieved bit error rate (BER). First, we examine the ability 9f 20-
of the noisy ICA algorithm to estimate the mixing matrix fx, 10

when used alone and its performance to approach to exact 4 0

blind detectors when it is used as a pre-processor device. The i, -10i


-1 -008 -0.6 -0.4 -02 0
Sampling Frequency f,
0.2 04 0.6 0Q8 1

system parameters are presented in table I.


Figure 3. Time and freq. characteristics of user 8's exact mixing matrix
columns.

241
Furthermore, the noisy ICA-MMSE detector tends to the exact demonstrates that switching to the noisy ICA post-processor is
blind MMSE detector as demonstrated by the time and interesting only if an improvement in the symbols estimation
frequency characteristics shown in Figs. 4 and 5 respectively. is intended, witch is observed at high SNR values
The same results were obtained when considering the noisy (SNR>15 dB). In the last test, the target user is set to a power
ICA-ZF and the exact ZF detectors which are practically 10 dB lower than the other active users and then its
indistinguishable. performance is compared to one of the users whose powers are
User IB's Bind ICA-MMSE Impulse Resp. - EaPy ponents equally distributed. The related system parameters are
0.21- 9 T -I
summarized in table IV. The BER plots of Fig. 8 show that the
02 inequality in power level that affects user 15 strongly
5 10 15
Chip Duration T:f/C
20 25 30 decreases the achieved performance of the linear receivers, the
Freq. Resp. of Bser
B's Blind ICA-MMEE Equalizer- EarIy Par blind ZF/Noisy ICA detector however performs comparable to
2
n
10
0
or even better than the blind ZF detector.
9
r
_10 _
TABLE II
2¢, -20~ -08 -06 -04 -02 0 02
Sampling Frequency f.1=/T
04 0.6 08 PARAMETERS SELECTION FOR A DOWNLINK MULTIUSER DS-CDMA SYSTEM
UNDER SLOW SELECTIVE FADING AND LOG-NORMAL POWER DISTRIBUTION
01 _ I~ , , ,
4, 0.05> + Parameter Value Parameter Value
PN codes type Large set # of users 10,20
-0w05 _ Kasami
10 15 20 25 Code length 63 _ Delays range (in chips) {0,..,31}
Chip Duration T/C Power distribution type Log-normal # of multipath rays 10
Number of simulated 1000 Std. deviations of multipath 0.3; 0.4
symbols
--7"' - - - and users' power dist.
I-.- -- - 11- . - --l.

-10
Achieved Bit Error with respect to SNP4
[ 20 10l
-1 -08 -06 -0A4 -02 0 0.2 04 0.6 08 iF --- Exact Blind ZF, K=10
Sampling Frequency f, I/T. - ZF/Noisy ICA, K=10
-- Exact Blind ZF; KPt20
- ZF/Noisy ICA, K 20
Figure 4. User 8's noisy ICA-MMSE detector impulse and frequency -

responses. 10-'

Uer S's EX4C0 Blind MMSE Impulst R4sp. -Early Cdrponehts


0,2 10-
02S :1 T
-0.2_
10 15 20 25 30 10
Chip Duration T
10 1. Resp. of User B's Exact Blind MMSE Equalizer- Erly Part

10
2 10 _
cr
E 0-220 08 -0 6 04 02 0 02 0.4 06 0.8
Sampling Frequency f.
llser B's Exact Blinid MM%SE Ir use Resp. - Ldfe Cronents 10 15 20 25
0.1F SNP (dB)

Figure 6. BER as a function of SNR with log-normal users' power


-0.1 tX distribution and short codes spreading.
0 10 15 20 25 30
Chip Duration T;

Freq. Pesp. of User B's Exact Blind MMS6 Equafizer- Late Part TABLE III
2 _ PARAMETERS SELECTION FOR A DOWNLINK MULTIUSER DS-CDMA SYSTEM
,! -lo _ UNDER SLOW SELECTIVE FADING, LARGE NETWORK SIZE, LOG-NORMAL
POWER DISTRIBUTION AND USING LONG CODES
2!
LL .2I
-213
08 0 04 -02 0 02 04 0= 6 08
Sampling Frequency fs Parameter Value Parameter Value
PN codes type Gold # of users 30,40
Figure 5. User 8's exact blind MMSE detector impulse and freq. responses. Code length
Power distribution type
Number of simulated
127
Log-nonnal
I
2000
Delays range (in chips)
of multipath rays

Std. deviations of multipath


I {0,..,63}
7
0.3; 0.4
Let's consider now various system configurations as symbols
--7 --- - - - - and users'
--- ---- power
F- - dist. -- --ll

illustrated in tables II-IV. Fig. 6 shows the BER performances


of the exact blind ZF detector compared to the blind ZF/Noisy TABLE IV
ICA detector according to the parameters of table II. The PARAMETERS SELECTION FOR A DOWNLINK MULTIUSER DS-CDMA SYSTEM
second schema provides a clear improvement in the achieved UNDER SLOW SELECTIVE FADING, CUSTOMIZED POWER DISTRIBUTION AND
USING LONG COMBINED CODES
BER and performs better that ZF detection at large SNR
Parameter Value Parameter Value
values even with log-normal power distribution and severe PN codes type Hadamard-Gold # of users 20
selective multipath channel. Fig. 7 illustrates the blind Code length 20x31=620 Delays range {0,..,309}
MMSE/Noisy ICA receiver performances compared with Power distribution type User 15 at -10 dB # of multipath rays 10
exact blind MMSE receiver for larger network size and using Number of simulated 1500 Std. deviations of 0.5; 0.5
symbols multipath and users'
longer codes as indicated in table III. This experience power dist.

242
Achieved Bit Error with respect to SNR [2] J. G. Proakis, Digital Communications, McGraw- Hill, New York, 4th
10
--- Exact Blind IIMMSE, K 30 ed., 2001.
MMSEfNoisy ICA K=30
-
-_- Exact Blind MMSEE K 40 [3] S. Verdu, "Minimum Probability of Error for Asynchronous Gaussian
MMSE/Noisy ICA,K 40 Multiple-Access Channels," IEEE Trans. on Information Theory, vol.
32, no. 1, pp. 85-96, Jan. 1986.
[4] M.L. Honig, U. Madhow and S. Verdu, "Blind Adaptive Multiuser
10' Detection," IEEE Trans. on Information Theory, vol. 41, no. 4, pp. 944-
960, Jul. 1995.
[5] X. Wang and H. V. Poor, "Blind Adaptive Interference Suppression for
0 10-
1i0 Direct-Sequence CDMA," Proceedings of the IEEE, vol. 86, no. 10, pp.
Lu i0-4 2049-2069, Oct. 1998.
[6] J. Joutsensalo and T. Ristaniemi, "Learning Algorithms for Blind
Multiuser Detection in CDMA Downlink," in Proc. IEEE 9th
International Symposium on Personal, Indoor and Mobile Radio
10-' Communication, London, U.K, vol. 3, pp. 1040-1044, Sep. 8-11, 1998.
[7] T. Ristaniemi and J. Joutsensalo, "Novel Schema for Blind Symbol
Separation in CDMA Downlink ," in Proc. 32th Asilomar Conference on
j0ol L
Signals, Systems and Computers, Monterey, Canada, vol.2, pp. 1853-
5NA (dB) 1857, Nov. 1-4, 1998.
[8] T. Ristaniemi and J. Joutsensalo, "Independent Component Analysis with
Code Information Utilization in DS-CDMA Signal Separation," in Proc.
Figure 7. BER as a function of SNR with log-normal users' power IEEE Global Telecommunications Conference (GLOBECOM'99), Rio
distribution, long codes spreading and large network size. de Janeiro, Brazil, vol. la, pp. 320-324, Dec. 5-9, 1999.
Achieved Bit Error with respect to SNR
[9] R. Cristescu, T. Ristaniemi, J. Joutsensalo and J. Karhunen, "Delay
10° Estimation in CDMA Communications Using a Fast ICA Algorithm," in
Proc. 2nd International Workshop on Independent Component Analysis
and Blind Source Separation (ICA'2000), Espoo, Finland, pp. 105-110,
10 Jun. 19-22, 2000.
l-,
[10] R. Cristescu, T. Ristaniemi, J. Joutsensalo and J. Karhunen, "CDMA
Delay Estimation Using Fast ICA Algorithm," in Proc. IEEE 11th
10' International Symposium on Personal, Indoor and Mobile Radio
Communications (PIMRC 2000), London, U.K, vol. 2, pp. 1117-1120,
Sep. 18-21, 2000.
103
[11] T. Ristaniemi and J. Joutsensalo, "Advanced ICA-Based Receivers for
DS-CDMA Systems," in Proc. IEEE ]]th International Symposium on
Personal, Indoor and Mobile Radio Communications (PIMRC2000),
10l
London, U.K, vol.1, pp. 276-28 1, Sep. 18-21, 2000.
-- Exact Blind ZF, User 8
[12] K. Waheed, K. Desai and F.M. Salem, "Natural Gradient Based Blind
ZF/Noisy ICA User 8
Exact Blind ZF, User 15
Multi User Detection in QPSK DS-CDMA Systems," in Proc. IEEE-
-
---
ZFlNoisy ICA User 15 INNS Joint International Conference On Neural Networks, Portland,
-25 -20 -10 -5 0 5 10 OR, vol. 3, pp. 1862-1867, Jul.. 20-24, 2003.
SNR (dB)
[13] 0. Ekici and A. Yongacoglu, "Application of Noisy-Independent
Figure 8. BER as a function of SNR with customized power distribution Component Analysis for CDMA Signal Separation," in Proc. IEEE 60th
and combined Hadamard-Gold spreading. Vehicular Technology Conference (VTC2004), Rhodes, Greece, vol. 5,
pp. 3812 - 3816, Sep. 26-29, 2004.
[14] K. Waheed and F. M. Salem, "Blind Information-Theoretic Multiuser
VI. CONCLUSIONS Detection Algorithms for DS-CDMA and WCDMA Downlink
Systems," IEEE Trans. on Neural Networks, vol. 16, no. 4, pp. 937-948,
In this paper, the efficiency of introducing ICA processing Jul. 2005.
in downlink DS-CDMA systems is tested. PCA decomposition [15] C. Jutten and J. Herault, "Blind Separation of Sources, Part I: An
is extremely important because it allows dimension reduction Adaptive Algorithm Based on Neuromimetic Architecture," Signal
which helps to reduce the additive noise, prevents over- Processing, vol. 24, pp. 1-10, 1991.
[16] P. Comon, "Independent Component Analysis- a New Concept?," Signal
learning and facilitates enormously the computations since Processing, vol. 36, pp. 287-314, 1994.
only 2K signals have to be estimated rather than C signals. [17] A. Hyvarinen and E. Oja, "Independent Component Analysis: Algorithms
Together with PCA, the noisy ICA algorithm may constitute and Applications," Neural Networks, 13(4-5), pp. 411-430, 2000.
an interesting and efficient tool for channel estimation and
[18] A. Hyvarinen and E. Oja, "A Fast Fixed-Point Algorithm for Independent
Component Analysis," Neural Computation, vol. 9, no. 7, pp. 1483-
symbols recovery when the number of users is relatively 1492, 1997.
small. Furthermore, the ICA post-processor shows in all [19] A. Hyvarinen, "Fast and Robust Fixed-Point Algorithms for Independent
studied cases performance that outperforms the exact blind Component Analysis," IEEE Trans. on Neural Networks, vol. 10, no. 3,
pp. 626-634, Apr. 1999.
detectors at relatively high SNR values even when the network [20] S. Bensley and B. Aazhang, "Subspace-Based Channel Estimation for
is characterized by severe fading multipath, long codes Code Division Multiple Access Communication Systems," IEEE Trans.
spreading, large number of users and dissimilarities in users' on Communications, vol. 44, no. 8, pp. 1009-1020, Aug. 1998.
power levels. [21] A. Hyvarinen, "Gaussian Moments for Noisy Independent Component
Analysis," IEEE Signal Processing Lett., vol. 6, no. 6, pp. 145-147, Jun.
1999.
REFERENCES [22] A. Hyvarinen, "Fast ICA for Noisy Data Using Gaussian Moments," in
Proc. IEEE International Symposium on Circuits and Systems
[1] R. Prasad, CDMAI- for Wireless Personal Communications, Artech House, (ISCAS'99), vol. 5, pp. 57-61, May 30-Jun. 2, 1999.
Boston, MA, pp.15-37, 1996.

243

Das könnte Ihnen auch gefallen